Study On Bearing Health Assessment Method Of Wind Turbine Gearbox Based On Deep Fusion Of Multi-domain Features

The bearings of wind turbine gearbox work under the severe working conditions of variable speed and variable load for a long time and their operation status directly affects the normal service of wind turbines.Simultaneously,wind farms usually need to purchase and prepare spare parts in advance to ensure normal and continuous production,due to the complicated processing and long procurement cycle of key components such as bearings of wind turbine gearbox.Therefore,a comprehensive and accurate assessment of the health status of the bearings of wind turbine gearbox can provide a reliable basis for formulating a reasonable and effective spare parts production plan and maintenance plan.On the premise of ensuring the safety and reliability of the unit,it can maximize the role of bearings and reduce Wind farm maintenance costs.The health assessment of the bearings of wind turbine gearbox is mainly to identify the current state of the bearings based on the monitoring signals and to predict their performance degradation trends and remaining useful life.At present,data-driven health assessment methods often use multi-domain feature fusion to more fully characterize the performance degradation process of bearings.Due to the complex and changeable operating conditions of the bearings of wind turbine gearbox,the degradation process is nonlinear.The traditional machine-learning model has a simple structure and poor ability to extract deep features,and cannot effectively fuse the degraded features of bearing.Deep learning has powerful deep feature extraction and non-linear fitting capabilities,showing great potential in the field of intelligent maintenance of wind turbine.Therefore,this paper uses deep learning to fuse multi-domain features to develop a health assessment method that focuses on the prediction of wind turbine gearbox bearing degradation trends and remaining life prediction.The main research contents of the paper are as follows:(1)When the life-cycle samples of wind turbine gearbox bearings are difficult to obtain,aiming at the problems of the nonlinear degradation process and difficult selection of degradation characteristics,a prediction method of wind turbine gearbox bearing degradation trend based on the Autoencoder(AE)and Gated Recurrent Unit(GRU)is proposed.At first,the multi-domain high-dimensional feature set of the bearing vibration signal is extracted,and the feature comprehensive evaluation value composed of the monotonic index and Spearman correlation coefficient is used to filter the features with better performance;then,the high-dimensional feature set of the AE fusion is used to obtain Non-linear characteristics.Finally,by utilizing the advantages of GRU neural network to efficiently process time-series signals,based on embedding theory and long-term iterative method to achieve the degradation trend prediction of wind power gearbox bearings,the future health status of bearings is mastered,with avoiding the risk of catastrophic damage to bearings.The validity and accuracy of the proposed method are proved by experimental verification and case analysis.(2)When the full life samples of wind turbine gearbox bearings are sufficient,the failure threshold is difficult to determine during prediction,and the multi-degradation features contribute differently to the model.A prediction method for the remaining life of the bearings of wind turbine gearbox with attention mechanism and multi-domain features is proposed.At first,the relative root mean square monitoring index is used to determine the initial recession point of the bearing,and the life cycle of the bearing is divided into two stages: the normal period and the recession period.Subsequently,the one-dimensional convolution neural network is used to automatically extract the degradation features of the bearing,and the attention mechanism combines high-dimensional abstract features to construct a virtual health indicator with a threshold to quantitatively evaluate the current health status of the bearing.Finally,through the prediction of the development trend of the health indicator based on the GRU neural network,obtained the remaining service life of the wind power gearbox bearing indirectly.Combining the experimental data of the bearing bench and the measured data of a wind field verifies the effectiveness of the method.(3)Based on the joint programming of C#/.Net general software development platform and Python/Keras deep learning framework,an assessment system for wind turbine gearbox bearing health was developed.The requirements analysis,overall design and functional module design were demonstrated.It realizes the compatibility of different platforms and programming frameworks,reliable data transmission,safe and efficient file classification management,and multi-threaded design.The system mainly includes basic signal analysis,initial recession point monitoring,health index construction,degradation trend prediction and remaining life prediction.The feasibility and practicability of the system are verified by the measured data of the bearings of the wind power gearbox.